Are you able to release the dataset so that we can reproduce your results? Otherwise, I’m unable to see the utility of your advice.
It is certainly possible that fork+marshal is faster for certain programs. But a generalised advice of “port all your ocaml 5 programs to fork+Marshall” is not useful. The advice has to be more specifc such as “If your program has a pattern X, then fork+marshal may offer better performance”. In your case, the work is embarrassingly parallel, and I assume fork+marshal works better (I presume).